Zhe Xie (谢哲)
PhD Candidate · Tsinghua University
About
I am a Ph.D. candidate in the Department of Computer Science and Technology at Tsinghua University, advised by Professor Dan Pei. My research focuses on multimodal LLM alignment and reasoning. I have end-to-end experience bringing new modalities such as time series into LLMs, spanning data construction, SFT alignment, RLVR, agentic RL, inference, and deployment.
Through ChatTS, I explored early directions in native time-series understanding and reasoning for multimodal LLMs. More recently, I have been working on LLM post-training across SFT, RL, and OPD/MOPD.
Prior to Tsinghua, I received my B.S. in Computer Science and Technology from Shanghai Jiao Tong University (GPA 3.98/4.3, Rank 4/147) in 2022, where I was a member of the Zhiyuan Honors Engineering Program (致远工科荣誉计划). I have held research internships at Moonshot AI, ByteDance, and eBay. Several algorithms from these experiences have been deployed in production systems.
News
Joined Moonshot AI (Kimi) as an RL Team Intern, working on LLM post-training.
One paper accepted to ICLR 2026
Paper accepted to ICSE 2026: "FoundRoot: Towards Foundation Model for Root Cause Analysis via Structured Deep Thinking."
Paper accepted to VLDB 2025: "ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced Understanding and Reasoning."
Joined ByteDance as a Research Intern, focusing on time series MLLMs.
Paper accepted to KDD 2024: "Microservice Root Cause Analysis with Limited Observability through Intervention Recognition in the Latent Space."
Selected Publications
† denotes equal contribution. For a complete list of publications, see my Google Scholar profile.
ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced Understanding and Reasoning
One of the first multimodal LLMs supporting time series as a numerical modality (TS-MLLM), enabling general understanding and reasoning over continuous numerical signals via SFT-based alignment rather than traditional forecasting or anomaly detection. Supports MTS, variable-length and complex numerical reasoning, outperforming text/image-based approaches such as GPT-4o. Built with controllable multimodal synthetic data, extending Qwen2.5/3 to incorporate numerical tokens, with a complete SFT and inference pipeline integrated into vLLM.
GitHub Stars: 460+; Hugging Face Stars: 140+; Citations: 110+; Model Downloads: 100,000+ (Jul. 2026).
FoundRoot: Towards Foundation Model for Root Cause Analysis via Structured Deep Thinking
We use RL to build one of the first LLM-based foundation models for root cause analysis, which has been deployed online.
Microservice Root Cause Analysis with Limited Observability through Intervention Recognition in the Latent Space
Multi-level root cause analysis under limited observability; algorithm deployed at eBay.
From Point-wise to Group-wise: A Fast and Accurate Microservice Trace Anomaly Detection Approach
First group-wise anomaly detection concept for traces; 20x speed improvement via graph algorithm.
Unsupervised Anomaly Detection on Microservice Traces through Graph VAE
Models traces as graphs for more accurate anomaly detection. Citations: 50+ (Mar. 2026).
Adversarial and Contrastive Variational Autoencoder for Sequential Recommendation
VAE for sequential recommendation. Citations: 140+ (Mar. 2026).
AutoDA-Timeseries: Automated Data Augmentation for Time Series
Experience
Research Internships
Post-Training
Contribute to Kimi across SFT, RL, and OPD/MOPD.
My work spans data and benchmark construction, training experiments, and evaluation.
Time Series Multimodal Large Language Models
· ChatTS: One of the first LLM-based foundation models for time series multimodal analysis.
· FoundRoot: One of the first RL-based LLM foundation models for root cause analysis.
· ThinkTime (under review): Achieving "Thinking with Time Series" with interleaved deep thinking of time series and Python tool use in LLM.
Corpus Construction and Fine-tuning for Customer Service Dialogue Models
Results deployed in production
AIOps
3 first-author CCF-A papers; results deployed in engineering
Education
Ph.D. in Computer Science and Technology
Advisor: Prof. Dan Pei · Research: Multimodal LLM, Anomaly Detection, Root Cause Analysis
B.S. in Computer Science and Technology
GPA: 3.98/4.3 · Rank: 4/147 · Zhiyuan Honors Engineering Program (致远工科荣誉计划)